fix(gpt-oss): emit fused raw expert tensors for SGLang

[Jiang020609]

Summary

Fixes #1840.

This fixes GPT-OSS raw Megatron-to-HF conversion for the non-colocate SGLang weight update path.

Previously, GPT-OSS expert tensors were emitted as per-expert gate_proj / up_proj / down_proj names. SGLang expects fused expert tensors for GPT-OSS, so the raw converter produced tensors that could not be loaded correctly by SGLang's fused MoE weight loader.

Changes

• Convert linear_fc1.weight into interleaved fused gate_up_proj.
• Transpose linear_fc2.weight before emitting down_proj.
• Fuse per-expert tensors into 3D tensors shaped [num_experts, ...] before returning them to the weight update path.
• Apply the same fused naming/layout for gate_up_proj_bias and down_proj_bias.
• Add CPU unit tests for GPT-OSS raw expert weight and bias conversion.

Validation

• python -m ruff check slime/backends/megatron_utils/megatron_to_hf/gpt_oss.py tests/test_gpt_oss_raw_converter.py
• python -m pytest tests/test_gpt_oss_raw_converter.py

GPU smoke on 1x NVIDIA A800-SXM4-80GB with lmsys/gpt-oss-20b-bf16:

• SGLang successfully loaded the GPT-OSS checkpoint.
• Prepared fused update tensors:
  • model.layers.0.mlp.experts.gate_up_proj: (32, 2880, 5760)
  • model.layers.0.mlp.experts.down_proj: (32, 2880, 2880)
  • model.layers.0.mlp.experts.gate_up_proj_bias: (32, 5760)
  • model.layers.0.mlp.experts.down_proj_bias: (32, 2880)
• engine.update_weights_from_tensor(...) returned (True, 'Success').
• Deterministic generation before and after the no-op raw expert update matched.
• Smoke result: GPT-OSS raw SGLang smoke passed.